Improving Predictive Accuracy in Elections.

The problem of accurately predicting vote counts in elections is considered in this article. Typically, small-sample polls are used to estimate or predict election outcomes. In this study, a machine-learning hybrid approach is proposed. This approach utilizes multiple sets of static data sources, such as voter registration data, and dynamic data sources, such as polls and donor data, to develop individualized voter scores for each member of the population. These voter scores are used to estimate expected vote counts under different turnout scenarios. The proposed technique has been tested with data collected during U.S. Senate and Louisiana gubernatorial elections. The predicted results (expected vote counts, predicted several days before the actual election) were accurate within 1%.